simonswiss/shot-doctor-models

Shot Doctor models

The pose models Shot Doctor runs in the browser to analyse basketball jump shots, hosted here so the app can load them from a CDN, pinned to a commit.

File	Model	Changes	Source	Licence
rtmw-m.onnx	RTMW-m whole-body pose: 133 keypoints (body, feet, face, hands), 256 × 192 input	Weights converted to fp16 (inputs and outputs stay fp32)	OpenMMLab RTMW, the rtmw-dw-l-m_simcc-cocktail14_270e-256x192 ONNX from rtmlib	Apache-2.0
pose_landmarker_full.task	MediaPipe Pose Landmarker (full)	None	Google MediaPipe	Apache-2.0

RTMW-m

• Input input: float32 [N, 3, 256, 192], RGB, normalised with mean (123.675, 116.28, 103.53) and std (58.395, 57.12, 57.375); the person's box padded 1.25× and widened or heightened to 3 : 4.
• Outputs simcc_x [N, 133, 384] and simcc_y [N, 133, 512]: the argmax of each over 2 gives the keypoint in input pixels, and the mean of the two maxima its score (COCO-WholeBody keypoint order).
• Converted with:

import onnx
from onnxconverter_common import float16

m = onnx.load('rtmw-dw-l-m_simcc-cocktail14_270e-256x192_20231122.onnx')
onnx.save(float16.convert_float_to_float16(m, keep_io_types=True), 'rtmw-m.onnx')

Runs on WebGPU in ONNX Runtime Web 1.30.